Audio large language models (Audio LLMs) exhibit systematic failures in transcribing code-switching speech despite strong multilingual capabilities. Focusing on English-Mandarin, we identify three failure modes: language omission, translation-instead-of-transcription, and hallucination. We apply Direct Preference Optimization (DPO) to align models, constructing preference pairs in which chosen responses preserve mixed-language content while rejected responses mimic failure patterns. Training three Audio LLMs on 100K pairs (570 hours), we observe consistent behavioral shifts: models learn to preserve language composition rather than translating when prompted for transcription. This alignment yields MER reductions up to 89.6% (in-distribution) and 20.0% (out-of-distribution). Our findings suggest DPO can effectively elicit correct code-switching transcription behavior from multilingual Audio LLMs.

翻译：音频大语言模型虽具备强大的多语言能力，但在转录语码转换语音时存在系统性缺陷。本研究聚焦英-中语码转换，识别出三大失效模式：语言缺失、翻译替代转录、以及内容幻觉。我们采用直接偏好优化方法对模型进行对齐，构建偏好对——其中选定响应保留混合语言内容，拒斥响应模仿上述失效模式。在三个音频大语言模型上使用10万对（570小时）训练数据后，观察到一致的行为转变：模型在收到转录指令时，能够学习保留语言构成而非进行翻译。该对齐方法使混合错误率降低高达89.6%（分布内测试）和20.0%（分布外测试）。研究结果表明，直接偏好优化能有效引导多语言音频大语言模型呈现正确的语码转换转录行为。